Presentation is loading. Please wait.

Presentation is loading. Please wait.

6.852: Distributed Algorithms Spring, 2008 Class 16.

Similar presentations


Presentation on theme: "6.852: Distributed Algorithms Spring, 2008 Class 16."— Presentation transcript:

1 6.852: Distributed Algorithms Spring, 2008 Class 16

2 Today’s plan Asynchronous shared-memory systems with failures. Consensus problem in asynchronous shared- memory systems. Impossibility of consensus [Fischer, Lynch, Paterson] Reading: Chapter 12 Next: Chapter 13

3 Asynchronous shared-memory systems with failures Process stopping failures. Architecture as for mutual exclusion. –Processes + shared variables, one system automaton. –Users Add “stop i ” inputs. –Effect is to disable all future non-input actions of process i. Fair executions: –Every process that doesn’t fail gets infinitely many turns to perform locally- controlled steps. –Just ordinary fairness---stop means that nothing further is enabled. –Users also get turns. p1p1 p2p2 pnpn x1x1 x2x2 A U1U1 U2U2 UnUn stop 1 stop 2 stop n

4 Consensus in Asynchronous Shared-Memory Systems Consensus in synchronous networks. –Algorithms for stopping failures: FloodSet, FloodMin, Optimizations: f+1 rounds, any number of processes, low communication –Lower bounds: f+1 rounds –Algorithms for Byzantine failures EIG: f+1 rounds, n > 3f, exponential communication –Lower bounds: f+1 rounds, n > 3f Asynchronous networks: Impossible Asynchronous shared memory: –Read/write variables: Impossible –Read-modify-write variables: Simple algorithms Impossibility results hold even if n is very large, f = 1.

5 Consequences of impossibility results Can’t solve problems like transaction commit, agreement on choice of leader, fault diagnosis,…in purely asynchronous model with failures. But these problems must be solved… Can strengthen the assumptions: –Timing assumptions: Upper and lower bounds on message delivery time, on step time. –Probabilistic assumptions And/or weaken the guarantees: –Small probability of violating safety properties, or of not terminating. –Conditional termination, based on stability for “sufficiently long” interval of time. We’ll see some of these strategies. But, first, the impossibility result…

6 Architecture V, set of consensus values Interaction between user U i and process (agent) p i : –User U i submits initial value v with init(v) i. –Process p i returns decision in decide(v) i. –I/O handled slightly differently from synchronous setting, where we assumed I and O in local variables. –Assume each user performs at most one init(v) i in an execution. Shared variable types: –Read/write registers (for now) p1p1 p2p2 pnpn x1x1 x2x2 A U1U1 U2U2 UnUn stop 1 stop 2 stop n pipi UiUi init(v) i decide(v) i

7 Problem requirements 1 Well-formedness: –At most one decide(*) i, appears, and only if there’s a previous init(*) i. Agreement: –All decision values are identical. Validity: –If all init actions that occur contain the same v, then that v is the only possible decision value. –Stronger version: Any decision value is an initial value. Termination: –Failure-free termination (most basic requirement): –In any fair failure-free (ff) execution in which init events occur on all “ports”, decide events occur on all ports. Basic problem requirements: Well-formedness, agreement, validity, failure-free termination.

8 Problem requirements 2: Fault-tolerance Failure-free termination: –In any fair failure-free (ff) execution in which init events occur on all ports, decide events occur on all ports. Wait-free termination (strongest condition): –In any fair execution in which init events occur on all ports, a decide event occurs on every port i for which no stop i occurs. –Similar to wait-free doorway in Lamport’s Bakery algorithm: says i finishes regardless of whether the other processes stop or not. Also consider tolerating limited number of failures. Should be easier to achieve, so impossibility results are stronger. f-failure termination, 0  f  n: –In any fair execution in which init events occur on all ports, if there are stop events on at most f ports, then a decide event occurs on every port i for which no stop i occurs. Wait-free termination = n-failure termination = (n-1)-failure termination. 1-failure termination: The interesting special case we will consider in our proof.

9 Impossibility of agreement Main Theorem [Fischer, Lynch, Paterson], [Loui, Abu-Amara]: – For n  2, there is no algorithm in the read/write shared memory model that solves the agreement problem and guarantees 1-failure termination. Simpler Theorem [Herlihy]: –For n  2, there is no algorithm in the read/write shared memory model that solves the agreement problem and guarantees wait-free termination. We’ll prove the simpler theorem first.

10 Restrictions (WLOG) V = { 0, 1 } Algorithms are deterministic: –Unique start state. –From any state, any process has  1 locally-controlled action enabled. –From any state, for any enabled action, there is exactly one new state. Non-halting: –Every non-failed process always has some locally- controlled action enabled, even after it decides.

11 Terminology Initialization: –Sequence of n init steps, one per port, in index order: init(v 1 ) 1, init(v 2 ) 2,…init(v n ) n Input-first execution: –Begins with an initialization. A finite execution  is: –0-valent, if 0 is the only decision value appearing in  or any extension of , and 0 actually does appear in  or some extension. –1-valent, if 1 is the only decision value appearing in  or any extension of , and 1 actually does appear in  or some extension. –Univalent, if  is 0-valent or 1-valent. –Bivalent, if each of 0, 1 occurs in some extension of .

12 Univalence and Bivalence univalent 0-valent  0 0 0 bivalent  0 1 1 1-valent  1 1 1

13 Exhaustive classification Lemma 1: –If A solves agreement with ff-termination, then each finite ff execution of A is either univalent or bivalent. Proof: –Can extend to a fair execution, in which everyone is required to decide.

14 Bivalent initialization From now on, fix A to be an algorithm solving agreement with (at least) 1-failure termination. –Could also satisfy stronger conditions, like f-failure termination, or wait-free termination. Lemma 2: A has a bivalent initialization. That is, the final decision value cannot always be determined from the inputs only. Contrast: In non-fault-tolerant case, final decision can be determined from the inputs only; e.g., take majority. Proof: –Same argument used (later) by [Aguilera, Toueg]. –Suppose not. Then all initializations are univalent. –Define initializations  0 = all 0s,  1 = all 1s. –  0 is 0-valent,  1 is 1-valent, by validity.

15 Bivalent initialization A solves agreement with 1-failure termination. Lemma 2: A has a bivalent initialization. Proof, cont’d –Construct chain of initializations, spanning from  0 to  1, each differing in the initial value of just one process. –Must be 2 consecutive initializations, say  and , where  is 0-valent and  is 1-valent. –Differ in initial value of some process i. –Consider a fair execution extending , in which i fails right after . –All but i must eventually decide, by 1-failure termination; since  is 0-valent, all must decide 0. –Extend  in the same way, all but i still decide 0, by indistinguishability. –Contradicts 1-valence of .  0 stop i all but i ’’ 0 stop i all but i

16 Impossibility for wait-free termination Simpler Theorem [Herlihy]: –For n  2, there is no algorithm in the read/write shared memory model that solves the agreement problem and guarantees wait-free termination. Proof: –We already assumed A solves agreement with 1-failure termination. –Now assume, for contradiction, that A (also) satisfies wait-free termination. –Proof based on pinpointing exactly how a decision gets determined, that is, how to move from bivalence to univalence.

17 Definition: A decider execution  is a finite, failure-free, input-first execution such that: –  is bivalent. –For every i, ext( ,i) is univalent. Lemma 3: A (with wait-free termination) has a decider execution. Impossibility for wait-free termination Extension of  with one step of i  1 n 2 bivalent univalent

18 Lemma 3: A (with w-f termination) has a decider. Proof: –Suppose not. Then any bivalent ff input-first execution has a 1-step bivalent ff extension. –Start with a bivalent initialization (Lemma 2), and produce an infinite ff execution  all of whose prefixes are bivalent. At each stage, start with a bivalent ff input-first execution and extend by one step to another bivalent ff execution. Possible by assumption. Impossibility for wait-free termination –  must contain infinitely many steps of some process, say i. –Claim i must decide in  : Add stop events for all processes that take only finitely many steps. Result is a fair execution . Wait-free termination says i must decide in .  is indistinguishable from , by i, so i must decide in  also. –Contradicts bivalence.  1 n 2 bivalent univalent

19 Impossibility for wait-free termination Proof of theorem, cont’d: –Fix a decider, . –Since  is bivalent and all 1-step extensions are univalent, there must be two processes, say i and j, leading to 0-valent and 1-valent states, respectively. –Case analysis yields a contradiction: 1. i’s step is a read 2. j’s step is a read 3. Both writes, to different variables. 4. Both writes, to the same variable.  1 n 2 bivalent univalent  i j 0-valent1-valent bivalent

20 Case 1: i’s step is a read Run all but i after ext( ,j). Looks like a fair execution in which i fails. So all others must decide; since ext( ,j), is 1-valent, they decide 1. Now run the same extension, starting with j’s step, after ext( ,i). They behave the same, decide 1. –Cannot see i’s read. Contradicts 0-valence of ext( ,i).  i j 0-valent1-valent bivalent all but i 1  i bivalent j all but i 1 0-valent

21 Case 2: j’s step is a read Symmetric. j  1-valent bivalent i 0-valent all but j 0  i bivalent 1-valenti 0-valent all but j 0 j

22 Case 3: Writes to different shared variables Then the two steps are completely independent. They could be performed in either order, and the result should be the same. ext( ,ij) and ext( ,ji) are indistinguishable to all processes, and end up in the same system state. But ext( ,ij) is 0-valent, since it extends the 0-valent execution ext( ,i). And ext( ,ji) is 1-valent, since it extends the 1-valent execution ext( ,j). Contradictory requirements. j  1-valent bivalent i 0-valent i j

23 Case 4: Writes to the same shared variable x. Run all but i after ext( ,j); they must decide. Since ext( ,j), is 1-valent, they decide 1. Run the same extension, starting with j’s step, after ext( ,i).  i j 0-valent1-valent bivalent all but i 1  i bivalent j all but i 1 0-valent They behave the same, decide 1. –Cannot see i’s write to x. –Because j’s write overwrites it. Contradicts 0-valence of ext( ,i).

24 Impossibility for wait-free termination So we have proved: Simpler Theorem: –For n  2, there is no algorithm in the read/write shared memory model that solves the agreement problem and guarantees wait-free termination.

25 Impossibility for 1-failure temination Q: Why doesn’t the previous proof yield impossibility for 1-failure termination? In proof of Lemma 3 (existence of decider), wait- free termination is used to say that a process i must decide in any fair execution in which i doesn’t fail. 1-failure termination makes a termination guarantee only when at most one process fails. Main Theorem: – For n  2, there is no algorithm in the read/write shared memory model that solves the agreement problem and guarantees 1-failure termination.

26 Impossibility for 1-failure temination From now on, assume A satisfies 1-failure termination, not necessarily wait-free termination (weaker requirement). Initialization lemma still works: –Lemma 2: A has a bivalent initialization. New key lemma, replacing Lemma 3: Lemma 4: If  is any bivalent, ff, input-first execution of A, and i is any process, then there is some ff-extension  of  such that ext( ,i) is bivalent. bivalent i  

27 Lemma 4  Main Theorem Lemma 4: If  is any bivalent, ff, input-first execution of A, and i is any process, then there is some ff-extension  of  such that ext( ,i) is bivalent. Proof of Main Theorem: –Construct a fair, ff, input-first execution in which no process ever decides, contradicting the basic ff-termination requirement. –Start with a bivalent initialization. –Then cycle through the processes round-robin: 1, 2, …, n, 1, 2, … –At each step, say for i, use Lemma 4 to extend the execution, including at least one step of i, while maintaining bivalence and avoiding failures. bivalent i  

28 Proof of Lemma 4 Lemma 4: If  is any bivalent, ff, input-first execution of A, and i is any process, then there is some ff-extension  of  such that ext( ,i) is bivalent. Proof: –By contradiction. Suppose there is some bivalent, ff, input-first execution  of A and some process i, such that for every ff extension  of , ext( ,i) is univalent. –In particular, ext( ,i) is univalent, WLOG 0-valent. –Since  is bivalent, there is some extension of  in which someone decides 1, WLOG failure-free. bivalent i   bivalent  i 0-valent

29 Proof of Lemma 4 –There is some ff-extension of  in which someone decides 1. –Consider letting i take one step at each point along the “spine”. –By assumption, results are all univalent. –0-valent at the beginning, 1- valent at the end. –So there are two consecutive results, one 0-valent and the other 1-valent: –A new kind of “decider”.  bivalent i 0-valent 1 i i i i i 1-valent univalent i i0-valent 1-valent j

30 New “Decider” Claim: j  i. Proof: –If j = i then: 1 step of i yields 0-valence 2 steps of i yield 1-valence –But process i is deterministic, so this can’t happen. “Child” of a 0-valent state can’t be 1-valent. The rest of the proof is a case analysis, as before… 0-valent 1-valent i i j

31 Case 1: i’s step is a read Run j after i. Executions ending with ji and ij are indistinguishable to everyone but i (because this is a read step of i). Run all processes except i in the same order after both ji and ij. In each case, they must decide, by 1-failure termination. After ji, they decide 1. After ij, they decide 0. But indistinguishable, contradiction! 0-valent 1-valent i i j j i i j 0-valent j 1-valent i i j 0-valent all but i 10

32 Case 2: j’s step is a read Executions ending with ji and i are indistinguishable to everyone but j (because this is a read step of j). Run all processes except j in the same order after ji and i. In each case, they must decide, by 1-failure termination. After ji, they decide 1. After i, they decide 0. But indistinguishable, contradiction! 0-valent 1-valent i i j 0-valent 1-valent i i j all but i 1 0

33 Case 3: Writes to different shared variables As for the wait-free case. The steps of i and j are independent, could be performed in either order, indistinguishable to everyone. But the execution ending with ji is 0-valent, whereas the execution ending with ij is 1-valent. Contradiction. i 1-valent j 0-valent j i

34 Case 4: Writes to the same shared variable x. As for Case 2. Executions ending with ji and i are indistinguishable to everyone but j (because the write step of j is overwritten by i). Run all processes except j in the same order after ji and i. After ji, they decide 1. After i, they decide 0. Indistinguishable, contradiction! 0-valent 1-valent i i j 0-valent 1-valent i i j all but i 1 0

35 Impossibility for 1-failure termination So we have proved: Main Theorem: –For n  2, there is no algorithm in the read/write shared memory model that solves the agreement problem and guarantees 1-failure termination.

36 Shared memory vs. networks Result also holds in asynchronous networks---revisit shortly. [Fischer, Lynch, Paterson 82, 85] proved for networks. [Loui, Abu-Amara 87] extended result and proof to shared memory.

37 Significance of [FLP 82, 85] For distributed computing practice: –Reaching agreement is sometimes important in practice: Agreeing on aircraft altimeter readings. Database transaction commit. –FLP shows limitations on the kind of algorithm one can look for. For distributed computing theory: –Variations: [Loui, Abu-Amara 87] Read/write shared memory. [Herlihy 91] Stronger fault-tolerance requirement (wait-free termination); simpler proof. –Circumventing the impossibility result: Strengthening the assumptions. Weakening the requirements/guarantees.

38 Strengthening the assumptions Using limited timing information [Dolev, Dwork, Stockmeyer 87]. –Bounds on message delays, processor step time. –Makes the model more like the synchronous model. Using randomness [Ben-Or 83][Rabin 83]. –Allow random choices in local transitions. –Weakens guarantees: Small probability of a wrong decision, or Small probability of not terminating, in any bounded time (Probability of not terminating approaches 0 as time approaches infinity.)

39 Weakening the requirements Agreement, validity must always hold. Termination required if system behavior “stabilizes”: –No new failures. –Timing (of process steps, messages) within “normal” bounds. Good solutions, both theoretically and in practice. [Dwork, Lynch, Stockmeyer 88]: Dijkstra Prize, 2007 –Keeps trying to choose a leader, who tries to coordinate agreement. –Coordination attempts can fail. –Once system stabilizes, unique leader is chosen, coordinates agreement. –Tricky part: Ensuring failed attempts don’t lead to inconsistent decisions. [Lamport 89] Paxos algorithm. –Improves on [DLS] by allowing more concurrency. –Refined, engineered for practical use. [Chandra, Hadzilacos, Toueg 96] Failure detectors (FDs) –Services that encapsulate use of time for detecting failures. –Develop similar algorithms using FDs. –Studied properties of FDs, identified weakest FD to solve consensus.

40 Extension to k-consensus At most k different decisions may occur overall. Solvable for k-1 process failures but not for k failures. –Algorithm for k-1 failures: [Chaudhuri 93]. –Impossibility result: [Herlihy, Shavit 93], [Borowsky, Gafni 93], [Saks, Zaharoglu 93] Godel Prize, 2004. Techniques from algebraic topology: Sperner’s Lemma. Similar to those used for lower bound on rounds for k- agreement, in synchronous model. Open question (currently active): –What is the weakest failure detector to solve k- consensus with k failures?

41 Importance of read/write data type Consensus impossibility result doesn’t hold for more powerful data types. Example: Read-modify-write shared memory –Very strong primitive. –In one step, can read variable, do local computation, and write back a value. –Easy algorithm: One shared variable x, value in V  {  }, initially . Each process i accesses x once. If it sees: – , then changes the value in x to its own initial value and decides on that value. –Some v in V, then decides on that value. Read/write registers are similar to asynchronous FIFO reliable channels---we’ll see the connection later.

42 Next time… Atomic objects Reading: Chapter 13


Download ppt "6.852: Distributed Algorithms Spring, 2008 Class 16."

Similar presentations


Ads by Google